Clique kernel support (#295)

* Adding experimental clique-based kernels (opt-in only)

Co-authored-by: Stanley Tsang <stanley.tsang@amd.com>
Co-authored-by: Gilbert Lee <gilbert.lee@amd.com>
Co-authored-by: Wenkai Du <43822138+wenkaidu@users.noreply.github.com>
Bu işleme şunda yer alıyor:
gilbertlee-amd
2020-11-10 15:44:10 -07:00
işlemeyi yapan: GitHub
ebeveyn 1fdb216f87
işleme 41bcfb8878
35 değiştirilmiş dosya ile 2204 ekleme ve 92 silme
Çalıştırılabilir dosya
+22
Dosyayı Görüntüle
@@ -0,0 +1,22 @@
#!/bin/bash
RCCL_INSTALL=../../build/release
EXE=$PWD/HelloRccl
LDPATH=$LD_LIBRARY_PATH:$RCCL_INSTALL
echo "Single process - With clique-based kernels:"
RCCL_CLIQUE_ALLREDUCE_BYTE_LIMIT=1073741824 RCCL_FORCE_ENABLE_CLIQUE=1 NCCL_DEBUG=INFO RCCL_ENABLE_CLIQUE=1 LD_LIBRARY_PATH=$LDPATH $EXE 4
echo "Single process - Without clique-based kernels:"
NCCL_DEBUG=INFO LD_LIBRARY_PATH=$LDPATH $EXE 4
echo "With clique-based kernels:"
RCCL_CLIQUE_ALLREDUCE_BYTE_LIMIT=1073741824 RCCL_FORCE_ENABLE_CLIQUE=1 NCCL_DEBUG=INFO RCCL_ENABLE_CLIQUE=1 NCCL_COMM_ID=$HOSTNAME:12345 LD_LIBRARY_PATH=$LDPATH $EXE 4 0 &
RCCL_CLIQUE_ALLREDUCE_BYTE_LIMIT=1073741824 RCCL_FORCE_ENABLE_CLIQUE=1 NCCL_DEBUG=INFO RCCL_ENABLE_CLIQUE=1 NCCL_COMM_ID=$HOSTNAME:12345 LD_LIBRARY_PATH=$LDPATH $EXE 4 1 &
RCCL_CLIQUE_ALLREDUCE_BYTE_LIMIT=1073741824 RCCL_FORCE_ENABLE_CLIQUE=1 NCCL_DEBUG=INFO RCCL_ENABLE_CLIQUE=1 NCCL_COMM_ID=$HOSTNAME:12345 LD_LIBRARY_PATH=$LDPATH $EXE 4 2 &
RCCL_CLIQUE_ALLREDUCE_BYTE_LIMIT=1073741824 RCCL_FORCE_ENABLE_CLIQUE=1 NCCL_DEBUG=INFO RCCL_ENABLE_CLIQUE=1 NCCL_COMM_ID=$HOSTNAME:12345 LD_LIBRARY_PATH=$LDPATH $EXE 4 3
echo "Without clique-based kernels:"
NCCL_COMM_ID=$HOSTNAME:12345 LD_LIBRARY_PATH=$LDPATH $EXE 4 0 &
NCCL_COMM_ID=$HOSTNAME:12345 LD_LIBRARY_PATH=$LDPATH $EXE 4 1 &
NCCL_COMM_ID=$HOSTNAME:12345 LD_LIBRARY_PATH=$LDPATH $EXE 4 2 &
NCCL_COMM_ID=$HOSTNAME:12345 LD_LIBRARY_PATH=$LDPATH $EXE 4 3